Development and validation of a pixel wise deep learning model to detect cataract on swept-source optical coherence tomography images.

Résumé

PURPOSE: The diagnosis of cataract is mostly clinical and there is a lack of objective and specific tool to detect and grade it automatically. The goal of this study was to develop and validate a deep learning model to detect and localize cataract on Swept Source Optical Coherance Tomography (SS-OCT) images.

METHODS: We trained a convolutional network to detect cataract at the pixel level from 504 SS-OCT images of clear lens and cataract patients. The model was then validated on 1326 different images of 114 patients. The output of the model is a map repreenting the probability of cataract for each pixel of the image. We calculated the Cataract Fraction (CF), defined as the number of pixel classified as "cataract" divided by the number of pixel representing the lens for each image. Receiver Operating Characteristic Curves were plotted. Area Under the Curve (ROC AUC) sensitivity and specitivity to detect cataract were calculated.

RESULTS: In the validsation set, mean CF was 0.024 ± 0.077 and 0.479 ± 0.230 (p < 0.001). ROC AUC was 0.98 with an optimal CF threshold of 0.14. Using that threshold, sensitivity and specificity to detect cataract were 94.4% and 94.7%, respectively.

CONCLUSION: We developed an automatic detection tool for cataract on SS-OCT images. Probability maps of cataract on the images provide an additional tool to help the physician in its diagnosis and surgical planning.